Efficient analytic continuation approach to Bethe-Salpeter excitation spectra in selected energy windows

This paper proposes an efficient analytic continuation method that constructs Bethe-Salpeter absorption spectra within specific energy windows by iteratively calculating polarizability tensors at a coarse set of complex frequencies to form a matrix-valued continued-fraction representation, which is then validated across diverse molecular and nanoscale systems.

The Gist

Imagine you are trying to listen to a specific song playing in a massive, noisy concert hall. The "song" is the way a molecule absorbs light (its spectrum), and the "noise" is the sheer number of tiny energy transitions happening inside the molecule.

Traditionally, to hear this song clearly, scientists have tried to find every single musician (every single energy state) in the orchestra, tune them one by one, and then figure out what the music sounds like. This is like trying to identify every single person in a stadium to understand the crowd's roar. It works, but it takes forever, especially if you only care about the high-pitched notes (high-energy X-rays) or the complex, buzzing sounds of a large crowd (plasmons).

This paper introduces a clever shortcut. Instead of listening to every single musician, the authors propose a method to guess the shape of the entire song by taking just a few strategic "sniff tests" in the air.

Here is how their approach works, broken down into simple concepts:

1. The "Sniff Test" (Sampling in the Complex Plane)

Imagine you want to know what a cake tastes like. Instead of baking the whole cake and eating it slice by slice, you dip a few toothpicks into the batter at specific spots.

• The Trick: The authors don't measure the light absorption at the "real" frequencies we see (like visible light colors). Instead, they take measurements at "imaginary" frequencies (a mathematical concept where the numbers have a "ghostly" imaginary part).
• The Result: They only need to take about 16 to 32 of these "sniff tests" (calculations) across a wide range of energies. This is much faster than calculating thousands of individual notes.

2. The "Magic Recipe" (Analytic Continuation)

Once they have these few data points, they use a mathematical tool called Analytic Continuation. Think of this as a master chef who, after tasting the batter at just a few points, can perfectly reconstruct the flavor of the entire cake, even the parts they didn't taste.

• They build a "continued fraction" (a specific type of mathematical recipe) that connects their few data points.
• This recipe allows them to predict exactly what the absorption spectrum looks like in the real world, right where we can measure it.

3. The "Group Portrait" vs. The "Individual Photos" (Tensor vs. Scalar)

This is a key innovation in the paper.

• The Old Way (Scalar): Imagine trying to reconstruct a 3D object by taking separate photos of its front, side, and back, and then trying to glue them together. Sometimes the pieces don't match up perfectly, and the picture looks blurry or distorted.
• The New Way (Tensorial): The authors treat the whole object as a single, unified 3D block. They calculate the "shape" of the entire object at once. This ensures that the "front," "side," and "back" all stay perfectly aligned.
• Why it matters: This makes the reconstruction much more stable and accurate, especially for complex molecules where the light interacts in many directions at once.

4. What They Found (The Results)

The authors tested this "shortcut" on several different "concerts":

• The Dipeptide (A small protein): They showed that their method could recreate the complex music of a small molecule using very few data points, whereas the old method would have needed to count hundreds of individual notes.
• The C60 Fullerene (A soccer-ball-shaped molecule): This molecule has a huge number of "dark" notes (sounds you can't hear) and only a few "bright" notes. Finding the bright notes the old way is like finding a needle in a haystack. Their method found the bright notes perfectly without needing to count the hay.
• The Silver Cluster (Ag20): This is a tiny metal ball that creates a "plasmon" (a collective wave of electrons). This isn't a single note; it's a massive, broad roar. Their method was perfect at capturing the envelope of this roar, smoothing out the chaos into a clear shape.
• X-Ray Absorption (Core Levels): Usually, to hear the high-pitched X-ray notes, you have to ignore all the low notes first (a process called CVS). The authors showed their method works just as well for these high notes without needing to throw away the low notes first, saving even more time.

The Bottom Line

The paper claims that you don't need to solve the entire puzzle to see the picture. By taking a few smart, strategic measurements in a mathematical "ghost" world, you can use a special recipe to reconstruct the full, real-world picture of how a molecule absorbs light.

The Catch:
Just like a recipe can only handle so many ingredients, this method has a limit. If a molecule has too many distinct, closely packed notes in a small range, the method might blend them together into one big blob. However, for most interesting cases—especially broad, complex sounds like plasmons or high-energy X-rays—it is a highly efficient and accurate way to get the job done.

Technical Summary

Technical Summary: Efficient Analytic Continuation Approach to Bethe-Salpeter Excitation Spectra

Problem Statement
The Bethe-Salpeter equation (BSE) formalism is a standard tool for calculating optical absorption spectra in solid-state physics and molecular chemistry, offering a valuable alternative to time-dependent Density Functional Theory (TD-DFT) by properly treating non-local electron-hole interactions. However, standard BSE calculations rely on solving the eigenvalue problem for the BSE Hamiltonian to obtain excitation energies and excitonic eigenstates. This approach faces significant computational bottlenecks:

1. Scaling: The search for all eigenstates scales as $O(N^6)$, while iterative methods (e.g., Davidson) targeting low-lying states scale as $O(N_{exc}N^4)$. Since the number of excitations ($N_{exc}$) in a given energy window scales quadratically with system size, iterative approaches become expensive as the spectrum densifies.
2. High-Energy/Deep Core Excitations: Calculating high-energy excitations (e.g., X-ray absorption) or core-to-valence transitions is particularly demanding because these states lie far above a vast number of valence excitations. While Core-Valence Separation (CVS) schemes can isolate core excitations, they still require converging a large number of states if the target window is broad.
3. Continuum/Resonance Structures: For systems with dense spectra (e.g., metallic clusters or plasmonic resonances), resolving individual eigenstates is computationally prohibitive and often unnecessary; the physical interest lies in the spectral envelope.
4. Limitations of Direct Resolvent Methods: Existing techniques that calculate the resolvent directly (e.g., via Haydock recursion or Lanczos) typically require converging all lower-energy states to access a specific high-energy window, making selective window targeting difficult.

Methodology
The authors propose an efficient approach to construct the BSE absorption spectrum in specific energy windows using analytic continuation (AC) techniques applied to the polarizability tensor.

• Core Concept: Instead of solving for eigenstates, the method calculates the BSE polarizability tensor $\bar{\bar{\alpha}}(z)$ iteratively at a coarse set of complex frequencies $\{z_k\}$ in the complex plane. These data points are used to construct a matrix-valued continued-fraction representation of the polarizability tensor, $\bar{\bar{\alpha}}^{AC}(z)$.
• Analytic Continuation Scheme: The authors utilize Thiele's interpolation formula to generate a rational function (continued fraction) that matches the reference tensor values at $\{z_k\}$. This functional form is then analytically continued close to the real energy axis to retrieve the absorption spectrum.
• Tensorial vs. Scalar Approach: A key methodological innovation is the construction of a single continued fraction for the full $3\times3$ polarizability tensor using matrix-valued coefficients, rather than independent scalar continuations for each tensor element.
  • This ensures that all matrix elements share the same pole structure (the eigenvalues of the Hamiltonian), avoiding inconsistencies found in scalar fits.
  • It allows capturing $n \times p$ poles (where $p$ is the tensor dimension) from $2n$ sampling points, providing a finer spectral structure than scalar approaches.
  • To handle numerical instabilities arising from rank-deficient matrices during recursion, the authors employ pseudo-inverses (via Singular Value Decomposition) to filter out negligible singular values.
• Sampling Strategy: The method employs a regular grid of complex frequencies $z_k = \omega_0 + k\Delta\omega + i\Gamma$. The imaginary part $\Gamma$ is chosen to be sufficiently large (e.g., 0.4–0.8 eV) to ensure stability, while the real part spans the target energy window. The authors also discuss exploiting symmetries (e.g., $f(z^*) = f(z)^*$) to double the effective number of sampling points without additional computational cost.
• Iterative Solver: The calculation of $\bar{\bar{\alpha}}(z_k)$ is performed by solving the linear system $(z\Delta - \Delta H_{BSE})X = D$ iteratively using the Generalized Minimal Residual (GMRES) method. An efficient preconditioner based on the Random Phase Approximation (RPA) susceptibility is introduced to accelerate convergence.
• Pole Extraction: The poles and residues of the continued fraction are extracted via linear algebra (solving a matrix pencil problem), allowing for the analysis of excitation energies and oscillator strengths directly from the continued fraction representation.

Key Contributions

1. Tensorial Continued Fraction: The paper demonstrates that building a matrix-valued continued fraction for the full polarizability tensor yields significantly more accurate and stable spectra compared to scalar approaches, particularly for systems with complex spectral features.
2. Selective Energy Window Access: The method allows for the calculation of absorption spectra in specific energy ranges (e.g., high-energy X-ray or specific valence windows) without needing to converge all lower-lying eigenstates, provided the sampling grid covers the target region.
3. Efficiency for Dense Spectra: The approach is shown to be highly efficient for capturing the "envelope" of broad structures composed of a quasi-continuum of excitations (e.g., surface plasmon resonances), where resolving individual states is computationally wasteful.
4. Core Level Applicability: The method is successfully applied to core-level excitations (X-ray Absorption Spectroscopy) in C60, demonstrating that it can treat deep core states with the same cost and accuracy as valence excitations, without strictly requiring CVS approximations (though CVS was tested and found robust).

Results
The authors validate the method on several systems:

• $\beta$-Dipeptide: A realistic test case showing that the tensorial approach captures the spectral weight distribution accurately with only 16 reference tensors, whereas scalar fits degrade at higher energies.
• C60 Fullerene (Valence): The method reproduces the main UV-Vis absorption peaks (3.8, 4.9, 6.0, 6.8 eV) with high accuracy. Using a coarse grid (AC(0.8)) yields errors of 20–40 meV for the main peaks, which improves to the meV level with a finer grid (AC(0.4)) or by enforcing conjugate symmetry. The method correctly identifies the degeneracy of bright excitations.
• PCBM Derivative: For the symmetry-broken PCBM molecule, where dark states become bright and the spectrum is dense, the AC method faithfully reproduces the broadened spectral envelope, capturing the complex structure where individual eigenstate resolution is difficult.
• Ag20 Silver Cluster: The method successfully captures the Localized Surface Plasmon Resonance (LSPR) envelope in the [3.9–4.1] eV range, demonstrating its utility for metallic clusters where excitations merge into a continuum.
• C60 Core Levels (XAS): The approach reconstructs the C1s X-ray absorption spectrum (284–293 eV) with good agreement to reference eigenstate calculations, confirming its applicability to high-energy core excitations.

Significance and Claims
The paper claims that this analytic continuation approach offers a computationally efficient alternative to standard eigenvalue solvers for BSE calculations, particularly when:

• Only specific energy windows are of interest.
• The system exhibits dense spectra or plasmonic resonances where an envelope description is sufficient.
• High-energy core excitations need to be accessed without the prohibitive cost of converging all intermediate valence states.

The authors emphasize that the method scales favorably (roughly $O(N^4)$ for the iterative steps) and that the tensorial matrix-valued formulation is crucial for stability and accuracy. They note that while the method merges closely lying peaks if the number of sampling points is insufficient, this can be mitigated by adaptive grid strategies. The work establishes a framework for calculating BSE spectra that bypasses the explicit diagonalization of the BSE Hamiltonian, making it a promising tool for large-scale systems and complex spectral regimes.